ORIGINAL ARTICLE
Paerns of olfactory dysfunction in chronic rhinosinusitis identified
by hierarchical cluster analysis and machine learning algorithms
Justin C. Morse, MD1 , Meghan H. Shilts, MS2 , Kim A. Ely, MD3 , Ping Li, MD1 , Quanhu Sheng, PhD4 ,
Li-Ching Huang, PhD4 , Todd J. Wannemuehler, MD1 , Naweed I. Chowdhury, MD1 ,
Rakesh K. Chandra, MD1 , Suman R. Das, PhD2 and Justin H. Turner, MD, PhD1
Background: Olfactory dysfunction is a common symp-
tom of chronic rhinosinusitis (CRS). We previously iden-
tified several cytokines potentially linked to smell loss,
potentially supporting an inflammatory etiology for CRS-
associated olfactory dysfunction. In the current study we
sought to validate paerns of olfactory dysfunction in CRS
using hierarchical cluster analysis, machine learning algo-
rithms, and multivariate regression.
Methods: CRS patients undergoing functional endoscopic
sinus surgery were administered the Smell Identification
Test (SIT) preoperatively. Mucus was collected from the
middle meatus using an absorbent polyurethane sponge
and 17 inflammatory mediators were assessed using a mul-
tiplexed flow-cytometric bead assay. Hierarchical cluster
analysis was performed to characterize inflammatory pat-
terns and their association with SIT scores. The random for-
est approach was used to identify cytokines predictive of
olfactory function.
Results: One hundred ten patients were enrolled in the
study. Hierarchical cluster analysis identified 5 distinct
CRS clusters with statistically significant differences in SIT
scores observed between individual clusters (p < 0.001).
A majority of anosmic patients were found in a single
O lfactory dysfunction is among the most common
symptoms of chronic rhinosinusitus (CRS) with a
1 Department of Otolaryngology–Head and Neck Surgery, Vanderbilt
University Medical Center, Nashville, TN; 2 Department of Medicine,
Vanderbilt University Medical Center, Nashville, TN; 3 Department of
Pathology, Microbiology, and Immunology, Vanderbilt University
Medical Center, Nashville, TN; 4 Department of Biostatistics, Vanderbilt
University Medical Center, Nashville, TN
Correspondence to: Justin H. Turner, MD, PhD, Department of
Otolaryngology–Head and Neck Surgery, Vanderbilt University Medical
Center, 1215 21st Avenue South, Suite 7209, Nashville, TN 37232-8605;
e-mail: justin.h.turner@vanderbilt.edu
Funding sources for the study: National Institutes of Health (RO3 DC014809
and L30 AI113795 to J.H.T.); National Center for Advancing Translational
Sciences (CTSA UL1TR000445); startup funds from the Vanderbilt University
Medical Center (P30 AI110527 and U19AI095227 to S.R.D.).
Disclosure: The content of this study is solely the responsibility of the
authors and does not necessarily represent official views of the National
Center for Advancing Translational Sciences or the National Institutes of
Health.
Potential conflict of interest: None provided.
cluster, which was additionally characterized by nasal poly-
posis (100%) and a high incidence of allergic fungal rhi-
nosinusitis (50%) and aspirin-exacerbated respiratory dis-
ease (AERD) (33%). A random forest approach identified a
strong association between olfaction and the cytokines in-
terleukin (IL)-5 and IL-13. Multivariate modeling identified
AERD, computed tomography (CT) score, and IL-2 as the
variables most predictive of olfactory function.
Conclusion: Olfactory dysfunction is associated with spe-
cific CRS endotypes characterized by severe nasal polypo-
sis, tissue eosinophilia, and AERD. Mucus IL-2 levels, CT
score, and AERD were independently associated with smell
loss. C 2018 ARS-AAOA, LLC.
Key Words:
anosmia; hyposmia; rhinosinusitis; endotype; mucus;
cytokine; interleukin; cluster analysis; machine learning
How to Cite this Article:
Morse JC, Shilts MH, Ely KA, et al. Paerns of ol-
factory dysfunction in chronic rhinosinusitis identified
by hierarchical cluster analysis and machine learning algo-
rithms. Int Forum Allergy Rhinol. 2018;00:1-10.
prevalence of between 30% and 80%.1, 2 Unfortunately,
the etiology of CRS-associated olfactory dysfunction re-
mains poorly understood. Olfactory loss in CRS has pre-
viously been attributed to an inability of odorants to ef-
fectively reach the olfactory cleft, due either to structural
abnormalities or presence of nasal polyps.3, 4 Recent re-
search has suggested that sinonasal inflammation may di-
rectly or indirectly affect olfactory neurons and olfactory
function.5 In animal models, certain cytokines have the abil-
ity to adversely affect olfactory neuron function, turnover,
and regeneration.5–8 An association between olfactory cleft
cytokine levels and olfactory function has been partially
Presented at American Rhinologic Society Fall 2018 Meeting, on October 6,
2018, in Atlanta, GA.
Received: 13 August 2018; Revised: 24 September 2018; Accepted:
28 October 2018
DOI: 10.1002/alr.22249
View this article online at wileyonlinelibrary.com.
International Forum of Allergy & Rhinology, Vol. 00, No. 0, xxxx 2018 1
Morse et al.
validated in human tissue by several groups.9, 10 Recently,
our group measured olfactory cleft mucus cytokine levels
in CRS patients and found that objective olfactory function
was inversely related to several cytokines, including inter-
leukin (IL)-2, IL-5, IL-6, IL-10, and IL-13. However, the
study was limited by a small sample size and was not able
to account for potential confounding factors.
In the current study we sought to validate patterns of
olfactory dysfunction in CRS using multiple complemen-
tary statistical approaches. We previously hypothesized
that mucus cytokine levels are reflective of olfactory in-
flammation and could be predictive of olfactory function.11
This study seeks to expand upon this hypothesis by incor-
porating inflammatory, clinical, and demographic factors,
with the ultimate goal of understanding constellations of
disease features associated with olfactory dysfunction.
Methods
Study design and population
This study was approved by the institutional review board
at Vanderbilt University. Patients presented to the Van-
derbilt Asthma, Sinus, and Allergy Program (ASAP) and
Otolaryngology Clinic at Vanderbilt’s Bill Wilkerson Cen-
ter. CRS was diagnosed according to the European Position
Paper on Rhinosinusitis and Nasal Polyps and the Interna-
tional Consensus Statement on Allergy and Rhinology, and
therefore were initially managed medically.12 Patients who
chose to undergo endoscopic sinus surgery were prospec-
tively enrolled. Only patients with diffuse, bilateral inflam-
matory CRS were included, and patients with odontogenic
rhinosinusitis, fungus balls, and isolated osteomeatal com-
plex obstruction were excluded. Patients were excluded
if they had received systemic steroids within 4 weeks of
surgery; had diagnosis of cystic fibrosis, autoimmune, or
granulomatous diseases; or were receiving immune-directed
monoclonal antibodies. Diagnosis of allergic rhinitis and
asthma was recorded. Allergic rhinitis was diagnosed based
on positive skin-prick testing and/or previous physician di-
agnosis and clinical history suggestive of seasonal variation
of atopic symptoms with improvement after use of topical
nasal steroid or oral antihistamines. Asthma was diagnosed
based on a positive methacholine challenge or consistent
pulmonary function studies, or by previous diagnosis by
a pulmonologist. All patients underwent a high-resolution
computed tomography (CT) scan of the paranasal sinuses
within 3 months of surgery. Each scan was evaluated by 2
physicians who were blinded to subject identifiers and diag-
nosis. A standard Lund-Mackay scoring system was used
to assess overall extent of CRS. Subjects enrolled in the
study also completed the 40-item Smell Identification Test
(SIT) immediately before surgery, which has been previ-
ously validated for olfaction assessment.13 Normative SIT
scores were extracted from the Smell Identification Test
Administration Manual (Sensonics International, Haddon
Heights, NJ). Raw scores were then adjusted for patient
2 International Forum of Allergy & Rhinology, Vol. 00, No. 0, xxxx 2018
age and gender by subtracting the mean normative age- and
sex-appropriate SIT score from the total SIT score for each
subject.10 A negative adjusted SIT score represents reduced
sense of smell compared with the mean for that subject’s
age and gender.
Mucus collection and histopathologic evaluation
of sinonasal tissue
At the beginning of surgery, 9 × 24-mm polyurethane
sponges (Summit Medical, St Paul, MN) were placed into
the middle meatus or ethmoid cavity of each subject un-
der endoscopic guidance as previously reported.11 This
approach has advantages over other methods for mucus
collection, such as standardization between subjects and
avoidance of specimen dilution. Each sponge was removed
after 5 minutes, placed in a sterile microcentrifuge tube, and
immediately processed. Sponges were placed into a micro-
porous centrifugal filter device (MilliporeSigma, Billerica,
MA) and centrifuged at 14,000g for 10 minutes to elute
mucus. Samples were then gently vortexed and again cen-
trifuged for 5 minutes to remove any cellular debris. Super-
natants were removed, placed into a new microcentrifuge
tube, and frozen at −80°C for later analysis. Cytokine as-
says were performed using a multiplex cytokine bead assay
(BD Biosciences, Franklin Lakes, NJ) according to the man-
ufacturer’s protocol, as described in previous studies.14, 15
Sinonasal tissue was collected from the ethmoid bulla or
ethmoid sinus in all patients undergoing endoscopic sinus
surgery for CRS. Eosinophil and neutrophil counts were
obtained from a dedicated, blinded histopathologic evalu-
ation of excised tissue by a pathologist and averaged over
5 randomly selected high-power fields.
Statistics
Sample size for principal component analysis and subse-
quent clustering was estimated by establishing a subject-
to-variable ratio of 5 (17 biologic variables, 110 subjects),
as recommended by O’Rourke and Hatcher and Gorsuch
et al.16, 17 Adequacy of the sample size was verified post hoc
by assessing variable communality (heavy loading of vari-
ables in retained components). Descriptive statistics and
frequency distributions were examined for each biologic
variable and all were positively skewed. To normalize data
for subsequent analysis, values were transformed by tak-
ing the square root, resulting in elimination or significant
reduction of skewing for all variables. A principal compo-
nent factor analysis with varimax rotation was then per-
formed on the transformed biologic variables. Variables
with a loading >0.5 were retained. The appropriate num-
ber of factors was selected by analysis of the Scree plot,
with a requirement that retained factors explain at least
70% of data variance, and that each factor has an eigen-
value >1.0. The regression method was then used to calcu-
late a factor score for each subject in each of the 5 factors.
Hierarchical cluster analysis was performed using Ward’s
method on squared Euclidian distances using the 5 factor
 Patterns of olfactory dysfunction in CRS

TABLE 1. Study population and demographics

All CRS CRSsNP CRSwNP p value

Number (%) 110 49 (45) 61 (55)

Age, years 48.15 ± 13.37 49.39 ± 13.86 47.15 ± 12.98 0.472
Sex, males, n (%) 59 (54) 24 (49) 35 (57) 0.443
Asthma, n (%) 46 (42) 12 (24) 34 (56) 0.007a
Allergic rhinitis, n (%) 74 (67) 27 (55) 47 (77) 0.024a
SNOT-22 score 44.0 (29.0-58.0) 48.0 (33.1-57.5) 43.0 (28.0-61.0) 0.678
CT score 15.0 (11.0-20.0) 12.0 (8.5-14.5) 18.0 (14.3-22.0) <0.001a
SIT score −7.0 (−2.0 to −23.1) −3.0 (−1.0 to −7.0) −20.0 (−5.0 to −26.5) <0.001a
Previous surgery, n (%) 41 (37) 12 (24) 29 (48) 0.017a
Tissue eos/HPF 25.7 (1.7-124) 2.0 (0.0-25.0) 80.6 (17.0-226.5) <0.001a
AERD 11 (10) 0 (0) 11 (20) <0.001a
AFRS 14 (13) 0 (0) 14 (25) <0.001a

*Data expressed as mean ± standard deviation or as median (interquartile range).

a
Statistically significant (p < 0.05).
AERD = aspirin-exacerbated respiratory disease; AFRS = allergic fungal rhinosinusitis; CRSsNP = chronic rhinosinusitis without nasal polyps; CRSwNP = chronic
rhinosinusitis with nasal polyps; CT = computed tomography; SD = standard deviation; SIT = Smell Identification Test; SNOT-22 = 22-item Sino-Nasal Outcome Test;
Tissue eos/HPF = tissue eosinophils per high-power field.

10 clusters and k was chosen at the break point where the

FIGURE 1. Dendrogram representing hierarchical cluster analysis of CRS
patients and relationship with olfactory function. Hierarchical cluster analysis
was performed using Ward’s method on squared Euclidian distances using
17 cytokines and inflammatory mediators as biologic variables. SIT score is
recorded in the right panel for individual study subjects and as a continuous
mean. Cluster 4, with the lowest SIT scores, is highlighted. CRS = chronic
rhinosinusitus; SIT = Smell Identification Test.
scores. The hierarchical structure of the data was visualized
using a dendogram. The appropriate number of clusters (k)
was selected using the elbow method. Total within sum
of squared error (SSE) was calculated for between 2 and
SSE started to smooth. Cluster stability was verified using
bootstrap analysis, with all clusters having a stability of at
least 0.7 (indicating plausible structure and good overall
cluster stability).18
Clusters were then retrospectively compared against the
individual components used for analysis, and then against
the individual biologic variables themselves. Subsequently,
clusters were compared against demographic and clinical
data. For comparison between groups, normality of data
was assessed using the D’Agostino-Pearson omnibus test.
Variables with a normal distribution were compared using
Student’s t test or analysis of variance, whereas nonpara-
metric data were analyzed using the Mann-Whitney test
or Kruskal-Wallis test. Comparative data were presented
as median with interquartile range. p < 0.05 was consid-
ered statistically significant for all comparisons. Statistical
analyses were performed with Prism 6 software (GraphPad,
Inc, La Jolla, CA), and principal component and hierarchi-
cal cluster analyses were performed using R version 3.4
(The R Project for Statistical Computing, Vienna, Austria;
http://www.R-project.org/).
The random forest algorithm was used to examine cy-
tokines that were most predictive of SIT score. Analysis
was performed in R using the randomForest package.19
The training and validation sets each represented half of
the samples, chosen at random without replacement. The
number of trees generated was 100, 1000, 10,000, and
1,000,000, with 5 variables (ie, cytokines) chosen at each
split. The percent variance explained appeared to level off
at 8% between 100,000 and 1,000,000 trees generated,

International Forum of Allergy & Rhinology, Vol. 00, No. 0, xxxx 2018 3
Morse et al.

TABLE 2. Characteristics of CRS clusters*

Cluster 1 (n = 46) Cluster 2 (n = 10) Cluster 3 (n = 37) Cluster 4 (n = 24) Cluster 5 (n = 27) p value

Age (years) 47.5 (41.3-57.5) 66.0 (56.8-70.1) 46.0 (39.0-58.0) 46.0 (38.5-53.0) 53.0 (39.0-60.5) 0.040a
Sex, n (% female) 23 (50) 5 (50) 22 (59) 8 (33) 6 (22) 0.030a
Race, n (% white) 39 (85) 9 (90) 33 (89) 19 (79) 24 (89) 0.970
Current smoker, n (%) 5 (11) 0 (0) 1 (3) 0 (0) 0 (0) 0.090
2
BMI (kg/m ) 28.7 (24.0-34.6) 28.3 (24.6-31.7) 30.6 (26.2-33.4) 29.0 (26.5-32.3) 29.3 (26.7-34.6) 0.910
Nasal polyps, n (%) 22 (48) 5 (50) 20 (54) 24 (100) 15 (56) <0.001a
Asthma, n (%) 14 (30) 4 (40) 16 (43) 19 (79) 24 (89) <0.001a
Allergic rhinitis, n (%) 30 (65) 5 (50) 21 (57) 17 (71) 20 (74) 0.490
AERD, n (%) 3 (7) 1 (10) 2 (5) 8 (33) 0 (0) <0.001a
AFRS, n (%) 3 (7) 0 (10) 1 (3) 12 (50) 2 (7) <0.001a
SNOT-22 score 45.6 ± 16.2 45.8 ± 29.0 41.9 ± 16.4 46.1 ± 22.1 50.2 ± 23.1 0.750
CT score 14.5 (11.1-17.8) 14.0 (11.0-18.5) 13.0 (11.0-16.0) 22.0 (20.0-23.0) 14.0 (11.0-16.0) <0.001a
SIT score −5.5 (−20.5 to −3.0) −7.0 (−7.0 to 0.5) −5.0 (−14.0 to 1.0) −25.0 (−29.0 to 20.0) −4.0 (−11.8 to −1.1) <0.001a
Previous surgery, n (%) 15 (33) 5 (50) 14 (38) 14 (58) 7 (26) 0.130
Tissue eos/HPF 27.5 (2.2-100.0) 10.0 (5.5-35.5) 30.0 (3.0-75.0) 91.5 (50.0-100.0) 18.0 (0.5-62.5) 0.001a
*
Data expressed as mean ± standard deviation or as median (interquartile range).
a
Statistically significant (p < 0.05).
AERD = aspirin exacerbated respiratory disease; AFRS = allergic fungal rhinosinusitis; BMI = body mass index; CT = computed tomography; SIT = Smell Identification
Test; SNOT-22 = Sino-Nasal Outcome Test; Tissue eos/HPF = tissue eosinophils per high-power field.

and therefore the number of trees was not further increased.
Variable importance plots were examined for both the
training and validation sets for each set of trees generated
to verify that the variable importance ordering remained
consistent.
Assessment of predictive variables on SIT scores, in-
cluding demographic characteristics and cytokines, was as-
sessed with univariate and multivariate regression model-
ing, performed using STATA release 15 (StataCorp LLC,
College Station, TX). In the multivariate model, all clinical
factors with p < 0.2 in the univariate modeling were in-
cluded in models for age- and sex-adjusted olfactory scores.
Collinearity diagnostics were performed using the variance
inflation factor and, when applicable, a correlation ma-
trix of the model was utilized to identify variables with
collinearity. If a variable was determined to be collinear, it
was dropped from the model and the model was reanalyzed
to determine effect on other predictor coefficients. p < 0.05
was considered statistically significant.
Results
Study population and demographics
Patients included in the study were undergoing func-
tional endoscopic sinus surgery for CRS and completed
the validated SIT immediately before their procedure.
A total of 110 patients with olfactory testing were en-
rolled, all of whom are part of an ongoing prospective
translational study, which has been partially character-
ized elsewhere.11, 14, 15 A majority of patients had nasal
polyps (55%), with comorbid asthma and allergic rhinitis
present in 42% and 67% of subjects, respectively (Table 1).
Eleven patients had aspirin-exacerbated respiratory disease
(AERD), whereas 14 were diagnosed with allergic fungal
rhinosinusitis (AFRS). Almost half of the subjects enrolled
had undergone previous endoscopic sinus surgery. Disease
burden was significant, with a median CT score of 15.0. The
median age- and sex-adjusted SIT score was −7.0 among
all CRS patients, with significant differences based on polyp
status. Patients with CRS with nasal polyps (CRSwNP) had
a median adjusted SIT score of −20.0 (−5.0 to −26.5)
compared with a median score of −3.0 (−1.0 to −7.0)
among those with CRS without nasal polyps (CRSsNP)
(p < 0.0001).
Olfactory function in inflammatory CRS clusters
In an earlier study we characterized several inflammatory
CRS endotypes using hierarchical cluster analysis of mu-
cus cytokines.14 In the process of validating these endo-
types, we repeated cluster analysis in an updated cohort
of 147 patients, 110 of whom had olfactory testing. Hi-
erarchical cluster analysis identified 5 CRS clusters with
unique inflammatory signatures (Fig. 1). Demographic and
clinical characteristics of each cluster are detailed in Ta-
ble 2. Age- and sex-adjusted SIT scores were significantly
different between clusters (p < 0.001) (Fig. 2A). Patients

4 International Forum of Allergy & Rhinology, Vol. 00, No. 0, xxxx 2018
 Patterns of olfactory dysfunction in CRS

FIGURE 2. SIT scores among inflammatory CRS clusters. (A) SIT scores
for individual patients in each cluster are presented as a scatterplot. Bars
represent the median and interquartile range. Cluster 4 demonstrates sig-
nificantly worse SIT scores compared with the other clusters (p < 0.001).
(B) Cluster 4 was associated with significantly higher prevalence of both
AERD and AFRS compared with other clusters (p < 0.001). AERD = aspirin-
exacerbated respiratory disease; AFRS = allergic fungal rhinosinusitis; CRS
= chronic rhinosinusitus; SIT = Smell Identification Test.

with the worst olfactory function were primarily concen-

trated within a single cluster (cluster 4) (Fig. 1). The me-
dian adjusted SIT score in this cluster was −25.0, indicative
of total anosmia (Fig. 2A).13 More than 80% of patients
in this cluster had either AERD or AFRS, both of which
varied significantly among clusters (p < 0.001) (Fig. 2B).
Cluster 4 was associated with a Th2-dominant signature,
with elevated levels of IL-5 (p < 0.001) and IL-13 (p <
0.001) compared with the other clusters (Fig. 3).
Use of machine learning algorithm to identify
cytokines impacting olfactory function
Hierarchical cluster analysis showed a close link between
poor olfactory function and a single inflammatory CRS
endotype, marked by elevated levels of T helper 2 (Th2)–
associated cytokines. This result is consistent with previous
reports that have associated decreased objective olfactory
function with Th2 cytokines, including IL-5 and IL-13.10,15
We sought to further validate these findings using a ran-
dom forest model, which is an ensemble machine learning
technique that fits decision trees using a random subset
of features to predict the outcome of interest. The relative
FIGURE 3. Cluster 4 is associated with elevated Th2 cytokines and anosmia.
Mucus cytokine levels for individual patients in each cluster are presented
as a scatterplot. Bars represent the median and interquartile range. Cluster
4 demonstrates significantly elevated IL-5 (A) (p < 0.001) and IL-13 (B) (p <
0.001) compared with the other clusters. IL = interleukin.
importance of predictors is then computed based on
the mean increase in error and decrease in node purity
when a variable of interest is excluded from the model.
Interestingly, this approach also identified IL-5 and IL-13
as the cytokines most predictive of olfactory function in
CRS (Fig. 4).19
Identification of variables affecting olfactory
function using multivariate regression
Our initial analysis and a small number of preceding
studies have identified eosinophilic inflammation and Th2
cytokines as potential mediators of olfactory dysfunction
in CRS.10, 11, 15 Small sample sizes have limited the ability

International Forum of Allergy & Rhinology, Vol. 00, No. 0, xxxx 2018 5
Morse et al.
FIGURE 4. Cytokines predictive of olfactory function using an ensemble learning method. Each cytokine or inflammatory mediator is ranked based on their
relative impact on decision tree construction. Results are presented as %IncMSE and IncNodePurity, both representative of the impacts of each variable on
the overall decision model. %IncMSE = increase in mean square error; IncNodePurity = increase in node purity.
of previous studies to account for covariates and other
potential confounding factors. We consequently incor-
porated a large number of demographic, clinical, and
inflammatory factors to further analyze CRS-associated
olfactory function in our large patient cohort. Univariate
regression identified asthma status (p = 0.016), polyp
status (p < 0.001), AERD (p < 0.001), CT score (p <
0.001), tissue eosinophilia (p = 0.002), and prior surgery
(p = 0.011) as variables predictive of olfactory function.
Cytokines associated with olfactory dysfunction included
IL-2 (p = 0.037), IL-5 (p = 0.001), and IL-13 (p < 0.001)
(Table 3). After multivariate analysis, only AERD
(p = 0.015), CT score (p = 0.014), and IL-2 (p = 0.005)
remained as predictive variables (Table 4). IL-5 and IL-13,
which were strongly associated with olfactory dysfunc-
tion after univariate analysis, demonstrated significant
collinearity, and this was verified using a correlation ma-
trix. Removal of either IL-5 or IL-13 from the model did not
significantly affect the results or the strength of the model.
Discussion
This study is the first to utilize hierarchal cluster analysis
and machine learning algorithms to assess the relationship
between inflammatory cytokines and CRS-associated
6 International Forum of Allergy & Rhinology, Vol. 00, No. 0, xxxx 2018
olfactory dysfunction. It is likewise the first study with an
adequate sample size for multivariate analysis of cytokines
and other potential contributors to olfactory loss. These
findings expand upon our group’s previous work and
potentially offer new insight into the potential role of
cytokine-associated inflammation in olfactory loss.11
Consistent with earlier studies,3, 20 our data suggest that
polyp status alone is not a sufficient predictor of olfactory
dysfunction. Most anosmic CRS patients were found in a
single CRS disease cluster that was chiefly characterized by
nasal polyps (100%); however, the majority of CRSwNP
patients did not appear in this cluster. Rather, characteris-
tics of this cluster were suggestive of a more severe form of
CRSwNP, with many patients diagnosed with either AERD
(33%) or AFRS (50%), and associated with a strong Th2-
dominant inflammatory signature. We previously showed
that olfactory cleft mucus cytokine levels correlate with
olfactory function in CRSwNP patients, and this was par-
ticularly true for the Th2-associated cytokines IL-5 and
IL-13.14 This association was also seen in our random
forest model, again suggesting that IL-5 and IL-13 were
the strongest predictors of olfactory dysfunction in CRS.
Surprisingly, our multivariate regression modeling did not
identify either cytokine as an independent predictor of ol-
factory dysfunction. This was largely due to collinearity of
 Patterns of olfactory dysfunction in CRS

TABLE 3. Olfactory assessment of all CRS patients both IL-5 and IL-13 with other variables in the model, sug-
(age/sex-adjusted SIT)—univariate regression analysis gesting that these cytokines may be markers of more severe
disease. This hypothesis was partially supported by a recent
Variables Unadjusted β 95% CI p value study in mice, which showed that allergic inflammation as-
Age −0.023 −0.180 to 0.135 0.770
sociated with elevated olfactory epithelium Th2 cytokines
reduced the number of immature olfactory neurons, but
Sex 0.422 −3.762 to 4.606 0.842 dis not affect the number of mature olfactory neurons or
Asthma −5.095 −9.213 to −0.978 0.016a olfactory function.21
Although our study did not confirm IL-5 and IL-13 as in-
Allergic rhinitis 0.909 −3.535 to 5.353 0.686
dependent effectors of olfactory loss in CRS, we did identify
AERD −15.162 −21.488 to −8.835 <0.001a a potential role for IL-2. Our previous study likewise identi-
Current smoker 3.943 −6.048 to 13.933 0.436 fied this cytokine as being closely correlated with olfactory
function.11 IL-2 is a nonspecific T-cell effector that regu-
NCS 4.504 −0.555 to 9.564 0.080
lates immunity and tolerance, but its role in CRS is poorly
Anti-leukotriene meds −4.020 −8.694 to 0.654 0.091 defined. A recent study showed that IL-2 may be associated
Previous surgery −5.470 −9.658 to −1.283 0.011a
with elevated immunoglobulin D levels and the presence of
pathogenic bacteria in CRSsNP patients.22 Potential func-
Eos/HPF (mean) −0.028 −0.045 to −0.010 0.002a tional relationships between IL-2 and the olfactory epithe-
Neu/HPF (mean) 0.033 −0.094 to 0.160 0.609 lium will require further investigation and confirmation.
Our study has identified AERD as being independently
Culture(+) purulence −1.589 −6.174 to 2.995 0.493
associated with olfactory dysfunction in CRS. The patho-
CT score −1.132 −1.477 to −0.787 <0.001a physiology of smell dysfunction in AERD is unclear, al-
Polyp status/phenotype −10.669 −14.342 to −6.997 <0.001a though recent studies have started to identify factors that
differentiate AERD from other CRSwNP patients.23 Both
AFRS −7.761 −13.660 to −1.863 0.010a
the innate and adaptive immune system have roles in AERD
IL-1β 0.001 −0.000 to 0.001 0.183 pathophysiology and severity.24 Both AERD and CRSwNP
IL-2 −0.012 −0.024 to −0.001 0.037a
are associated with eosinophilic tissue inflammation, yet
studies have generally not shown significant differences in
IL-4 −0.433 −1.212 to 0.345 0.272 the number of tissue eosinophils in each group. Conversely,
IL-5 −0.013 −0.020 to −0.005 0.001a the eosinophil degranulation product, eosinophil cationic
protein (ECP), is elevated in AERD patients compared with
IL-6 0.000 −0.001 to 0.000 0.887
CRSwNP patients. This suggests that eosinophils may be
IL-7 −0.003 −0.095 to 0.089 0.950 more highly activated in AERD.25, 26 Elevation of Th2-
IL-8 0.000 −5.170e-06 to 0.000 0.245 associated cytokines has been reported in both CRSwNP
and AERD, yet a specific difference in inflammatory signa-
IL-9 −0.031 −0.104 to 0.041 0.396
tures has not been clearly defined.27 Of note, hierarchical
IL-10 0.002 −0.016 to 0.019 0.858 cluster analysis from this and previous studies by our group
IL-12/I-23p40 0.003 −0.005 to 0.011 0.464
do suggest that AERD may be associated with a specific
inflammatory CRS endotype.14 The relationship between
IL-13 −0.036 −0.051 to −0.020 <0.001 AERD and olfaction is even less clear. Gudziol et al demon-
IL-17A −0.209 −0.562 to 0.145 0.244 strated that AERD patients had worse olfaction at baseline,
which subsequently improved after aspirin desensitization;
IL-21 0.002 −0.008 to 0.012 0.701
however, no current studies, to our knowledge, have evalu-
TNF-α 0.005 −0.014 to 0.024 0.627 ated inflammatory profiles and olfaction in these patients.23
IFN-γ 0.031 −0.038 to 0.101 0.374 Furthermore, smell testing has not been found to be predic-
tive of AERD.23 Multiple studies have demonstrated greater
Eotaxin −0.030 −0.062 to 0.002 0.068
disease severity among patients with AERD, based on both
RANTES 0.000 −0.000 to 0.001 0.400 endoscopy28 and CT scores.29 This suggests that reduced
olfactory identification scores in AERD may be multifac-
a
Statistically significant (p < 0.05).
AERD = aspirin-exacerbated respiratory disease; AFRS = allergic fungal rhi- torial, and likely due to collective differences in disease
nosinusitus; CI = confidence interval; CRS = chronic rhinosinusitis; CT = com- severity, polyp burden, and inflammatory signatures.30
puted tomography; Eos/HPF = eosinophils per high-power field; IL = interleukin;
INF = interferon; NCS = nasal corticosteroid medication; Neu/HPF = neutrophils Eosinophilic inflammation in allergic mouse models
per high-power field; RANTES = regulated-on-activation, normal T-cell expressed has previously been shown to have adverse effects on
and secreted; TNF = tumor necrosis factor.
the olfactory epithelium,31 and human studies suggested
similar findings.10, 32 Although the exact mechanisms of
eosinophil-associated olfactory loss remains unclear, it
is well established that eosinophilia is correlated with

International Forum of Allergy & Rhinology, Vol. 00, No. 0, xxxx 2018 7
Morse et al.

TABLE 4. Objective olfactory assessment of all CRS patients (age/sex-adjusted SIT)—multivariate gression analysis

Collinearity statistics

Variables Adjusted β 95% CI p value VIF

Age Not modeled – – –

Sex Not modeled – – –
Asthma 0.726 −3.366 to 4.818 0.725 1.47
Allergic rhinitis Not modeled – – –
AERD −9.102 −16.385 to −1.820 0.015 a
1.79
Current smoker Not modeled – – –
NCS use 1.942 −2.522 to 6.336 0.39 1.15
Anti-leukotriene meds 1.913 −2.511 to 6.336 0.393 1.31
Previous surgery −1.623 −5.464 to 2.218 0.404 1.24
Eos/HPF (mean) −0.007 −0.024 to 0.010 0.418 1.44
Neu/HPF (mean) Not modeled – – –
Culture(+) purulence Not modeled – – –
CT score −0.569 −1.019 to −0.118 0.014a 1.93
Polyp status/phenotype −4.123 −8.844 to 0.598 0.086 1.98
AFRS −1.959 −8.319 to 4.402 0.542 1.58
IL-1β Not modeled – – –
IL-2 −0.014 −0.024 to −0.004 0.005 a
1.11
IL-4 Not modeled – – –
IL-5 −0.004 −0.017 to 0.009 0.547 4.51
IL-6 Not modeled – – –
IL-7 Not modeled – – –
IL-8 Not modeled – – –
IL-9 Not modeled – – –
IL-10 Not modeled – – –
IL-12/I-23p40 Not modeled – – –
IL-13 0.004 −0.028 to 0.036 0.802 5.62
IL-17A Not modeled – – –
IL-21 Not modeled – – –
TNF-α Not modeled – – –
IFN-γ Not modeled – – –
Eotaxin −0.010 −0.040 to 0.021 0.525 1.38
RANTES Not modeled – – –
a
Statistically significant (p < 0.05).
AERD = aspirin-exacerbated respiratory disease; AFRS = allergic fungal rhinosinusitus; CI = confidence interval; CRS = chronic rhinosinusitis; CT = computed tomography;
Eos/HPF = eosinophils per high-power field; IL = interleukin; INF = interferon; NCS = nasal corticosteroid medication; Neu/HPF = neutrophils per high-power field;
RANTES = regulated-on-activation, normal T-cell expressed and secreted; TNF = tumor necrosis factor; VIF = variance inflation factor.

8 International Forum of Allergy & Rhinology, Vol. 00, No. 0, xxxx 2018
 Patterns of olfactory dysfunction in CRS

a Th2 inflammatory profile.33 Both local neurotoxicity
secondary to release of eosinophilic granule proteins34
and eosinophil-associated cytokine effects35 have been
postulated as possible mechanisms of eosinophil-associated
olfactory loss. Interestingly, although tissue eosinophilia
was associated with olfactory loss in our univariate model,
the multivariate analysis failed to support this link. Rather,
our data suggest that eosinophilia may instead simply be in-
dicative of more severe disease, with olfactory dysfunction
being one of many indicators of disease severity.
Our study does have some limitations that should be ac-
knowledged. First, we assessed smell function using the
semiobjective SIT, rather than using formal and more
quantitative assessment tools. Although the SIT is a well-
established method for assessment of smell function that is
highly correlated with threshold testing, it remains possi-
ble that some differences in olfactory function could have
been overlooked in this study. This possibility is partially
supported by a small number of recent studies. For exam-
ple, Lavin et al found that Charcot-Leyden crystal protein
gene expression in superior turbinate tissue was associated
with olfactory thresholds, but not olfactory identification.32
Conversely, Kohli et al found that elevated olfactory cleft
IL-5 levels were associated with worse identification scores,
but did not affect thresholds or discrimination.2 The rela-
tionships identified in the current study will ultimately need
to be validated using objective and quantitative olfactory
testing. Second, it is possible that mucus cytokine levels
may show temporal variations, particularly in relation to
to CRS and comorbid disease exacerbations. Although we
have attempted to limit the impact of this potential prob-
lem by assessing olfactory function and cytokine levels on
the same day, subsequent studies that assess temporal vari-
ations in individual cytokines and any potential effects on
olfactory function may help to clarify this issue.
To our knowledge, our study is the largest to date to
evaluate potential associations between sinonasal inflam-
mation and olfaction in CRS patients. Strengths of this
study include its prospective design, evaluation of a wide
array of cytokines and inflammatory mediators, and use of
multiple complementary statistical approaches. This find-
ings continue to underscore the limitations of phenotypic
categorization of CRS, and further suggest that olfactory
loss may be more closely associated with endotypic, rather
than phenotypic differences.
Conclusion
Anosmia in CRS is associated with a Th2-driven inflam-
matory CRS endotype enriched in AERD and AFRS pa-
tients. Previously reported associations between the Th2-
associated cytokines IL-5 and IL-13 and olfactory function
were not confirmed after multivariate analysis, whereas IL-
2 was the only cytokine independently associated with smell
dysfunction in CRS. In addition, certain disease characteris-
tics, including radiographic severity and presence of AERD,
were also independently associated with olfactory dysfunc-
tion. These results suggest that a combination of inflamma-
tory, clinical, and demographic factors likely contribute to
olfactory loss in CRS patients.

References
1. Jiang R-S, Lu F-J, Liang K-L, et al. Olfac-
tory function in patients with chronic rhinosinusi-
tis before and after functional endoscopic sinus
surgery. Am J Rhinol. 2008;22:445-448. https://doi.
org/10.2500/ajr.2008.22.3195.
2. Kohli P, Naik AN, Harruff EE, Nguyen SA,
Schlosser RJ, Soler ZM. The prevalence of
olfactory dysfunction in chronic rhinosinusitis.
Laryngoscope. 2017;127:309-320. https://doi.org/
10.1002/lary.26316.
3. Doty RL, Mishra A. Olfaction and its alter-
ation by nasal obstruction, rhinitis, and rhinosinusi-
tis. Laryngoscope. 2001;111:409-423. https://doi.org/
10.1097/00005537-200103000-00008.
4. Nishijima H, Kondo K, Yamamoto T, et al. Influence
of the location of nasal polyps on olfactory airflow and
olfaction. Int Forum Allergy Rhinol. 2018;8:695-706.
https://doi.org/10.1002/alr.22089.
5. Sousa Garcia D, Chen M, Smith AK, Lazarini PR,
Lane AP. Role of the type I tumor necrosis fac-
tor receptor in inflammation-associated olfactory dys-
function. Int Forum Allergy Rhinol. 2017;7:160-168.
https://doi.org/10.1002/alr.21855.
6. Turner JH, Liang KL, May L, Lane AP. Tumor necro-
sis factor alpha inhibits olfactory regeneration in a
transgenic model of chronic rhinosinusitis–associated
olfactory loss. Am J Rhinol Allergy. 2010;24:336-340.
https://doi.org/10.2500/ajra.2010.24.3498.
7. Lane AP, Turner J, May L, Reed R. A genetic
model of chronic rhinosinusitis-associated olfac-
tory inflammation reveals reversible functional im-
pairment and dramatic neuroepithelial reorganiza-
tion. J Neurosci. 2010;30:2324-2329. https://doi.
org/10.1523/JNEUROSCI.4507-09.2010.
8. Turner JH, May L, Reed RR, Lane AP. Reversible
loss of neuronal marker protein expression in a trans-
genic mouse model for sinusitis-associated olfactory
dysfunction. Am J Rhinol Allergy. 2010;24:192-196.
https://doi.org/10.2500/ajra.2010.24.3460.
9. Oyer SL, Mulligan JK, Psaltis AJ, Henriquez OA,
Schlosser RJ. Cytokine correlation between sinus tis-
sue and nasal secretions among chronic rhinosinusi-
tis and controls. Laryngoscope. 2013;123:E72-E78.
https://doi.org/10.1002/lary.24305.
10. Hauser LJ, Chandra RK, Li P, Turner JH. Role of
tissue eosinophils in chronic rhinosinusitis–associated
olfactory loss. Int Forum Allergy Rhinol. 2017;7:957-
962. https://doi.org/10.1002/alr.21994.
11. Wu J, Chandra RK, Li P, Hull BP, Turner JH.
Olfactory and middle meatal cytokine levels corre-
late with olfactory function in chronic rhinosinusi-
tis. Laryngoscope. 2018;128:E304-E310. https://doi.
org/10.1002/lary.27112.
12. Orlandi RR, Kingdom TT, Hwang PH. Interna-
tional Consensus Statement on Allergy and Rhi-
nology: Rhinosinusitis Executive Summary. Int Fo-
rum Allergy Rhinol. 2016;6:S3-S21. https://doi.org/
10.1002/alr.21694.
13. Doty RL, Shaman P, Kimmelman CP, Dann M.
University of Pennsylvania Smell Identification Test:
a rapid quantitative olfactory function test for
the clinic. Laryngoscope. 1984;94:176-178. http://
www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retr
ieve&db=PubMed&dopt=Citation&list_uids=
6694486.
14. Turner JH, Chandra RK, Li P, Bonnet K, Schlundt
DG. Identification of clinically relevant chronic rhi-
nosinusitis endotypes using cluster analysis of mucus
cytokines. J Allergy Clin Immunol. 2018;141:1895-
1897.e7. https://doi.org/10.1016/j.jaci.2018.02.002.
15. Turner JH, Li P, Chandra RK. Mucus T
helper 2 biomarkers predict chronic rhinosinusi-
tis disease severity and prior surgical interven-
tion. Int Forum Allergy Rhinol. 2018;8:1175-1183.
https://doi.org/10.1002/alr.22160.
16. O’Rourke N, Hatcher L. A step-by-step approach to
using SAS for factor analysis and structural equa-
tion modeling’ 2nd ed. 2013, SAS Institute Inc.,
Cary, North Carolina. Anal Struct Model. 2013:9.
https://doi.org/10.1002/9781118411360.wbcla114.
17. Richard L. Gorsuch. Factor Analysis. 1983. 2nd ed.
Erlbaum. Hillsdale, New Jersey.
18. Hennig C. Cluster-wise assessment of cluster sta-
bility. Comput Stat Data Anal. 2007;52:258-271.
https://doi.org/10.1016/j.csda.2006.11.025.
19. Breiman L. Random forests. Mach Learn. 2001;45:5-
32. https://doi.org/10.1023/A:1010933404324.
20. Hox V, Callebaut I, Bobic S, Jorissen M, Hellings
PW. Nasal obstruction and smell impairment in
nasal polyp disease: correlation between objective and
subjective parameters. Rhinology. 2010;48:426-432.
https://doi.org/10.4193/Rhino10.049.
21. Rouyar A, Classe M, Gorski R, et al. Type 2/Th2-
driven inflammation impairs olfactory sensory neuro-
genesis in mouse chronic rhinosinusitis model. Allergy.
2018. https://doi.org/10.1111/all.13559.
22. Min J-Y, Nayak JV, Hulse KE, et al. Evidence
for altered levels of IgD in the nasal airway
mucosa of patients with chronic rhinosinusitis.
J Allergy Clin Immunol. 2017;140:1562-1571.e5.
https://doi.org/10.1016/j.jaci.2017.05.032.
23. Gudziol V, Michel M, Sonnefeld C, Koschel D, Hum-
mel T. Olfaction and sinonasal symptoms in patients
with CRSwNP and AERD and without AERD: a
cross-sectional and longitudinal study. Eur Arch
Otorhinolaryngol. 2017;274:1487-1493. https://
doi.org/10.1007/s00405-016-4366-x.
24. Laidlaw TM, Boyce JA. Aspirin-exacerbated res-
piratory disease—new prime suspects. New Engl
J Med. 2016;374:484-488. https://doi.org/10.1056/
NEJMcibr1514013.
25. Stevens WW, Schleimer RP. Aspirin-exacerbated res-
piratory disease as an endotype of chronic rhinosinusi-
tis. Immunol Allergy Clin North Am. 2016;36:669-
680. https://doi.org/10.1016/j.iac.2016.06.004.

International Forum of Allergy & Rhinology, Vol. 00, No. 0, xxxx 2018 9
Morse et al.
26. Han JK. Subclassification of chronic rhinosi-
nusitis. Laryngoscope. 2013;123(Suppl 2):S15-S27.
https://doi.org/10.1002/lary.23979.
27. Van Zele T, Claeys S, Gevaert P, et al. Dif-
ferentiation of chronic sinus diseases by mea-
surement of inflammatory mediators. Allergy Eur
J Allergy Clin Immunol. 2006;61:1280-1289.
https://doi.org/10.1111/j.1398-9995.2006.01225.x.
28. Jang DW, Comer BT, Lachanas VA, Kountakis SE.
Aspirin sensitivity does not compromise quality-of-
life outcomes in patients with Samter’s triad. Laryn-
goscope. 2014;124:34-37. https://doi.org/10.1002/
lary.24220.
29. Robinson JL, Griest S, James KE, Smith TL. Impact
of aspirin intolerance on outcomes of sinus surgery.
Laryngoscope. 2007;117:825-830. https://doi.org/10.
1097/MLG.0b013e3180333121.
10 International Forum of Allergy & Rhinology, Vol. 00, No. 0, xxxx 2018
30. Bochenek G, Kuschill-Dziurda J, Szafraniec K,
Plutecka H, Szczeklik A, Nizankowska-Mogilnicka
E. Certain subphenotypes of aspirin-exacerbated res-
piratory disease distinguished by latent class anal-
ysis. J Allergy Clin Immunol. 2014;133:98-103.e6.
https://doi.org/10.1016/j.jaci.2013.07.004.
31. Epstein VA, Bryce PJ, Conley DB, Kern RC,
Robinson AM. Intranasal Aspergillus fumigatus
exposure induces eosinophilic inflammation and
olfactory sensory neuron cell death in mice.
Otolaryngol Head Neck Surg. 2008;138:334-339.
https://doi.org/10.1016/j.otohns.2007.11.029.
32. Lavin J, Min JY, Lidder AK, et al. Supe-
rior turbinate eosinophilia correlates with ol-
factory deficit in chronic rhinosinusitis patients.
Laryngoscope. 2017;127:2210-2218. https://doi.org/
10.1002/lary.26555.
33. Vlaminck S, Vauterin T, Hellings PW, et al. The im-
portance of local eosinophilia in the surgical outcome
of chronic rhinosinusitis: a 3-year prospective observa-
tional study. Am J Rhinol Allergy. 2014;28:260-264.
https://doi.org/10.2500/ajra.2014.28.4024.
34. Acharya KR, Ackerman SJ. Eosinophil gran-
ule proteins: Form and function. J Biol Chem.
2014;289:17406-17415. https://doi.org/10.1074/jbc.
R113.546218.
35. Li L, Walker TL, Zhang Y, Mackay EW,
Bartlett PF. Endogenous interferon gamma di-
rectly regulates neural precursors in the non-
inflammatory brain. J Neurosci. 2010;30:9038-9050.
https://doi.org/10.1523/JNEUROSCI.5691-09.2010.